Ranking methodology

BetGlitch does not train a predictive model. Generation 4 combines a specialist provider's fixture signal, its per-league model report card, its dedicated value-bet model and a verified bookmaker price. This page states the eligibility rule exactly and names the version that produced any given commitment.

The purpose is not to make the process look sophisticated. It is to make every decision interpretable: what data was available, which rule ranked it, what price was recorded, what was missing and what the result later showed.

Current ranking version

rank-v4-a323d53a53716d00

This string is stamped onto every snapshot and every public commitment, and it sits inside the commitment's integrity hash — so it cannot be altered after the fact. It is derived from the parameters below, not typed by hand: change any of them and the version changes with them.

Earlier commitments can carry the legacy label consensus_ensemble. That label is preserved rather than rewritten. New commitments use the derived version above.

Why a version exists at all

BetGlitch expects this logic to change — that is the point of measuring it in public. But if the rule changed silently, the growing record would blend several different systems into one average that describes none of them. Stamping the version means any future performance figure can be split by the exact rule that produced it, and a change of rule can never be hidden inside an aggregate.

Parameters in force

Markets eligible for value selection
1x2
Confidence floor (leading outcome probability)
0.55
Minimum gap to next outcome — draw
0.1
Minimum gap to next outcome — other
0.08
Provider fixture must be predictable
yes
Rejected league predictability labels
poor, low, bad
Rejected predictive-power trend
down
Active, outcome-aligned provider value bet
required
Correct-score agreement
supporting ranking evidence
Double-chance support
supporting ranking evidence
Maximum explicit cross-market contradictions
1
Minimum canonical-price buffer above fair odd
2%
Minimum bookmakers quoting selection
3
Maximum relative price spread
15%
Maximum recorded price age
12 hours
Maximum selections per run
3
Gem probability source
price-implied baseline probability (1 / price)
Gem payout source
verified canonical multi-bookmaker quote
Gem primary ranking
kelly-style probability-payout balance (ranking only, not a stake)
Selection rule
probability-payout balance, supporting cross-market evidence, league predictability, predictive-power trend, price advantage, model probability, hit ratio, then price dispersion

How qualified Gems are ordered

Ranking happens only after every eligibility gate above has passed. The provider's baseline price supplies the external implied chance (p = 1 / baseline price), while the canonical quote supplies the verified payout. The primary balance is (p × verified price − 1) / (verified price − 1).

Correct-score and double-chance agreement strengthen this ordering, but neither is an independent veto because both are correlated with the same underlying fixture evidence. Two explicit contradictions still reject a Gem. This prevents a large price or a large price difference from winning the ranking automatically. It is a conservative ordering device, not a BetGlitch-calibrated probability, a stake size or betting advice. Provider quality, trend, hit ratio and price dispersion are deterministic tie-breakers.

What these parameters do not give you

  • Provider agreement and a quoted edge are not a guarantee. A value candidate can lose, and a short run can be extremely poor.
  • This exact generation has not yet accumulated enough settled, out-of-sample decisions to establish profitability. Its results must be judged separately from the earlier score-ranking method. The measured evidence is published and recomputed from live data.
  • Other markets remain in the research evidence feed, but they are not promoted into betting selections until a market-specific value signal and enough forward evidence exist.
  • The provider's fair odd and league report card are external model outputs, not BetGlitch guarantees. The public results are still the final test of whether this filtering works in practice.

How improvement is judged

  1. 1. Preserve: freeze decisions under a named version before kickoff.
  2. 2. Measure: publish eligible wins and losses, score separation and pricing evidence.
  3. 3. Change: alter a rule only as a new version, never inside an existing record.
  4. 4. Compare: judge versions separately before claiming that one improved the evidence.

Results →External timestamps →Measured separation →Probability calibration →